The idea moved beyond a camera log
The first version recognized a meal, remembered an explicit promise from a conversation, and suggested an evening walk. I then extended it into a context orchestrator. During a user-approved session, LifeLens can combine vision, voice, maps, movement, weather, calendar, and a previously authorized health profile to decide what might be useful now.
A meal can become a calorie range and an optional running-time comparison based on a saved pace and prior InBody record. A run can become a rain warning and a well-lit route home. A conversation can become a commitment the user may add to a calendar. Each result is a proposal; it is not an automatic command.
What I actually developed
The public prototype includes an operable React experience, first-person demo scenes, an animated glasses HUD, a FastAPI service, a Strands agent backed by Amazon Bedrock, typed Pydantic contracts, deterministic fixtures, and safety tests. The backend rejects retained raw media and refuses actions that do not carry explicit confirmation.
The hardware boundary is deliberately replaceable. Vuzix M400/M4000 has a native Android prototype with Camera2, temporary speech, D-pad controls, and explicit approval. The repository also defines a phone-local perception interface and a shared MomentEvent schema. Meta's official Wearables Device Access Toolkit supplies the Ray-Ban Meta camera stream and Mock Device Kit for iOS and Android companion apps. Both paths minimize frames locally before optional cloud reasoning.
Permission is not one big switch
Camera access does not imply microphone access. Location access does not imply location sharing. A health record does not become permanent agent memory. LifeLens scopes each signal independently for the active session, transforms raw input into a minimal structured observation, and discards the media. Messages, reminders, route changes, logs, and stored memories remain separate confirmation steps.
Conversation analysis is restricted to observable behavior and explicit statements. The agent may remember “I will send the slides Friday,” but it may not decide that the other person dislikes the user, infer personality, or suggest manipulation. Food and exercise numbers stay estimates and are framed as planning comparisons, never punishment or medical advice.
What remains to become a real device
The hosted experience uses simulated wearable events; the agent and consent boundary are real, but there is no claim that continuous everyday capture has been solved. A hardware pilot still needs battery and thermal testing, background-audio restrictions, bystander notice, regional recording-law review, safe fallback behavior, and evaluation of whether interventions are actually helpful instead of merely frequent.